I wonder if this has to do with integrating into modeling/vfx software like Houdini and Blender. Those “AI” tools are actually pretty good use cases for the tech.
I wonder if this has to do with integrating into modeling/vfx software like Houdini and Blender.
How’s the energy demands and carbon output of that compared to prior conventional methods? I don’t know; I’m actually asking.
If it’s anything like the giant coal-powered hell factory databases that tech startups are using and expanding right now to chase the “AI” hype dragon, we don’t need more of that.
Depends on texture size and model. The models before LLMs and SDF hype are extremely efficient with the right libraries. Without getting to into the weeds these older ones are no different then doing any other image processing (like applying a gauss blur on 8k image or something). Newer stuff uses diffusion models. So something like a 500x500 texture probably is the equivalent to leaving 25 watt lightbulb on for an hour at worst. But you do it once and it repeats so it’s efficient. I know the early models were trained on cropped animal patterns, zoomed in materials, and other very vanilla datasets.
No problem. It’s appalling companys are pushing this out at scale and virtually uncapped. These models use as much power as large scale scientific simulations/models. Video more so. You don’t want a lot of people running these concurrently. VERY BAD lol
I wonder if this has to do with integrating into modeling/vfx software like Houdini and Blender. Those “AI” tools are actually pretty good use cases for the tech.
Edit: I mean the texture generators specifically.
How’s the energy demands and carbon output of that compared to prior conventional methods? I don’t know; I’m actually asking.
If it’s anything like the giant coal-powered hell factory databases that tech startups are using and expanding right now to chase the “AI” hype dragon, we don’t need more of that.
Depends on texture size and model. The models before LLMs and SDF hype are extremely efficient with the right libraries. Without getting to into the weeds these older ones are no different then doing any other image processing (like applying a gauss blur on 8k image or something). Newer stuff uses diffusion models. So something like a 500x500 texture probably is the equivalent to leaving 25 watt lightbulb on for an hour at worst. But you do it once and it repeats so it’s efficient. I know the early models were trained on cropped animal patterns, zoomed in materials, and other very vanilla datasets.
I appreciate the information.
No problem. It’s appalling companys are pushing this out at scale and virtually uncapped. These models use as much power as large scale scientific simulations/models. Video more so. You don’t want a lot of people running these concurrently. VERY BAD lol